Recent applications of packet based IP switching technology have extended to many areas, such as mobile infrastructure, Multi-Service Provisioning Platform (MSPP), high-speed Internet routers, Storage Area Network (SAN) equipment, and high-definition television. The current demand for higher bandwidth across the global network is driving a need for higher performance and higher port counts in “next generation” switching solutions.
Many of these high-speed switches are built around a crossbar architecture because of its speed and simplicity. Due to switching speed limitations, crossbar based architectures typically use input queues to hold packets or cells waiting to be transferred. A typical switching scheme applies the well known first in/first out (FIFO) regime to prioritizing the transfers from these queues. However, such simple FIFO input queues inherently suffer diminished performance because of head of line (HOL) blocking.
HOL blocking can be eliminated by applying a virtual output queue (VOQ) FIFO structure, but this poses strenuous demands on memory size. One such demand is that a VOQ FIFO architecture requires a memory size to grow according to the square of the number of port increases. Another such memory demand is presented by the per-flow queuing for quality of service (QOS).
The provision of QOS demanded by modern, advanced-architecture network systems requires the isolation of specific traffic flows among multiple flows, and each specific flow needs to have its own behavior characteristics and service level. A prerequisite for such isolation typically requires separate FIFO implementation for each flow. Together, the need for separate VOQ FIFOs for each QOS priority can drive a switch fabric solution to hundreds of chips. Therefore, providing enough memory within the switching function becomes a costly bottleneck for next generation switch fabric chip set vendors.
This can be illustrated by the following example. For a 32-port switch fabric with 32 priorities, 1024 FIFOs (32 destinations times 32 priorities) are required per port for per-flow QOS support and for elimination of HOL blocking. If the fixed cell size of 64 bytes is used and the cell depth of each VOQ is 64, then the overall memory requirement for the switch fabric system is 128 Mbytes (1024 VOQs times 64 cells per queue times 64 bytes per cell times 32 ports).
This huge memory size would be extraordinarily difficult and expensive, and perhaps impossible with contemporary processing technologies, to integrate into a cost effective, single chip solution. Thus, another approach should be taken to achieve small die size but without sacrificing performance. Examining actual internet traffic indicates that average usage of this huge input buffer can be as low as less than 2% at 99% full input traffic rates. This low actual memory utilization rate offers opportunity for a different approach. However, taking advantage of the low actual memory utilization rate is problematic using conventional crossbar architectures.
Conventional crossbar architectures utilize separated queues for the detection and scheduling of cell transfer. This is problematic because of the large capacity, speed, and addressability demands made on memory resources and/or the large memory size such an approach requires, as discussed above. Further, even conventional crossbar switch architectures made with memories of capacity and addressability sufficiently large to accommodate such demands can pose other complicating problems.
Even for conventional crossbar switch architectures made with memories of capacity and addressability sufficiently large to accommodate the demands of detection and scheduling of cell transfer, implementing a FIFO regime thereon requires a pointer functionality to nominate and designate cells for transfer. The pointer functionality requires a pointer as well as a management system and/or process for its control. Implementing a FIFO regime with a large memory in a crossbar switch demands a complex pointer management system, which is difficult and expensive to implement in conventional technology.
A further problem with conventionally implementing crossbar switch architectures made with memories of capacity and addressability sufficiently large to accommodate the demands of detection and scheduling of cell transfer is that of retarded speed. The switching speeds of such conventional crossbar switch architectures are constrained by the addressability and size of the memory they field.
Multicasting can conserve bandwidth because the same data is sent to multiple destinations, which can improve switching efficiencies. However, multicasting requires the movement of data from a single input port to as many output ports as there are destinations. In conventional crossbar switches, the problems of memory size and resource utilization, pointer complexity, and speed can be exacerbated by the additional demands unique to switching data from a single input port to a multiplicity of output ports.
Also, in conventional crossbar switches, prioritizing data to be multicasted with respect to data that is to be unicasted can be problematic, and prioritization problems can further exacerbate the problems described above. Further, in conventional crossbar switches, prioritizing data to be multicasted with respect to other data also to be multicasted, especially amongst yet other data that is to be unicasted can make the problem even more complicated. Further still, some conventional crossbar switches may delay all unicast traffic while multicast data is being serviced, which can reduce efficiency.
The conventional art is problematic therefore because memories of sufficient capacity and addressability for implementing a FIFO switching function in a crossbar switch using conventional architectures therefore are difficult and expensive to achieve, especially to support more than one QoS level, and in a single integrated circuit (IC; e.g., chip). The conventional art is also problematic because even if an adequate memory is achieved, the FIFO switching function of such a crossbar switch requires a complex pointer management system. Further, the conventional art is problematic because crossbar switches so constructed may operate at less than optimal switching speeds and bandwidths. These problems characterizing conventional crossbar switches are exacerbated when data is to be multicasted. Further, multicasting adds the problems of prioritizing multicast data with respect to unicast data as well as to other multicast data and can delay unicasting.